crungus amongus

About

Crungus Amongus combines a comparative study of model outputs with an interactive audiovisual archive. It investigates how consistently different models respond to prompts containing “crungus”, and what similarities appear across their outputs. Shared prompts make those differences available for inspection.

Browse the images by model, explore visual similarities in the atlas, or examine variation with the eigencrungi sliders. The mosaic fills a screen with images drawn at random, and crungus radio plays independently shuffled sounds over it.

The word

In June 2022, comedian Guy Kelly shared a grid of nine DALL·E mini images generated from the word “crungus”. Several resembled a hairy, horned creature. The images became known as a digital cryptid. This project uses that recurring resemblance as a starting point for comparing models.

The archive

Models are drawn from Replicate's text-to-image collection and a curated selection of older models. This is a sample of available endpoints, including related models and fine-tunes. It is not an exhaustive survey or a set of independent model families.

The image experiment requests 100 outputs per model for each of two prompts: “crungus” and “a picture of a crungus”. We supply no stylistic instructions or fixed seeds. Endpoint defaults and required inputs vary between models. Where an endpoint exposes safety controls, we use the most permissive settings available. As of 2026-10-17, the archive holds 24,400 images from 122 models.

Timeline years identify model releases, using curated dates where available and Replicate version dates as a fallback. The images were generated for this project; they are not historical samples collected in those release years. The homepage filmstrip selects the image with the highest average CLIP similarity to the other images in each release-year group.

The sound

The audio experiment requests ten outputs for each of “crungus” and “the sound of a crungus”. Models come from Replicate's music collection and a curated selection including Bark. We request twenty-second clips where duration is configurable; other endpoints set their own length. Required lyrics use “crungus”, repeated where necessary to meet input constraints. Clips are encoded and loudness-normalised for playback. The archive holds 310 clips from 16 models.

Consistency

Consistency is the mean pairwise cosine similarity between a model's output embeddings for one prompt. Images use CLIP embeddings; audio uses CLAP embeddings. A high score can reflect recurring subjects, composition or style. It does not establish that a model represents a particular creature. Even a model that always produces a beige void could score highly.

Archive cards show the higher of a model's two prompt scores; model pages show each prompt separately. The scores describe similarity within each model's outputs, not agreement between models or output quality. Image and audio scores use different embedding models and should be compared within their respective collections.

Refusals

Early runs used each endpoint's default safety settings. The refusal figures on model pages record those runs, measured before the configuration changed on 9 September 2026. Subsequent runs use the most permissive safety settings each endpoint exposes; provider restrictions may still apply. The historical refusal rates do not describe the images currently displayed.

The atlas

The atlas projects CLIP image embeddings into two dimensions using UMAP. Nearby images can share subjects, colours or composition. Apparent clusters and the gaps between them depend on the projection: UMAP can introduce separations. The map supports visual exploration; it does not establish distinct categories of creature.

The eigencrungi

The pixel sliders combine a mean image with principal components fitted to the archive. Images are centre-cropped and resized to 128 × 128 pixels, without aligning subjects. The components capture variation in colour and framing as well as depicted subjects. Combining the retained components gives an approximation of an archive image.

The CLIP features mode applies principal component analysis to ViT-bigG-14 image embeddings, each containing 1,280 values. Kandinsky 2.2 renders selected positions along individual components. These previews are the decoder's interpretations; compare them with the nearby archive images shown alongside. Kandinsky is also one of the models represented in the collection.

Stability measures how closely a component reproduces across disjoint halves of the archive, averaged over eight splits. We use a cosine similarity threshold of 0.7 when naming components and rendering their filmstrips. This is a working threshold, not a boundary between meaningful features and noise. Current scores appear beside the sliders.

The separate “crungus-ness” slider uses image similarity to CLIP's text embedding of “crungus”. It measures an association in CLIP, distinct from the within-model consistency score. Its pixel visualisation is fitted from the archive by least squares, so that visualisation depends on the sampled images.

Technical notes document earlier sample-size experiments, component rotation, decoder configuration and generation costs. Those historical observations are kept separate from the current component scores.

Prior sightings

The data

The archive is also published as static JSON, readable from any origin. /api/models.json lists the models, /api/images.json and /api/audio.json list every image and clip, and /api/models/<slug>.json holds one model's outputs. Each entry is a link to the file with the model and prompt that produced it. To sample an output, fetch a list and select an entry.

Colophon

images and clips generated via the Replicate API · scored with open-clip (ViT-B/32) and LAION CLAP · projected with UMAP · decomposed with numpy · encoded as AVIF and Opus (with an AAC fallback) · served from Tigris object storage · built with Astro · made by Ben Swift at the ANU School of Cybernetics